![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F799299%2Ff4a1323e0364e1b75d9da7f818a92265%2FScreen%20Shot%202022-08-04%20at%209.36.10%20AM.png?generation=1659620213373671&alt=media\")
Hi,
\n
\n\nQualitatively, what is the difference between CA and LAA? I cannot tell by just looking at the data.
\n
From what I've gleaned from the literature, it's tough for even trained professionals to discriminate CE/LAA. I think part of this competition is to figure out if computational techniques can do any better.
\n\n\nThere are some examples where there are multiple separate pieces. An example is 2b49d6_0. Can we say that each piece would show the label or some pieces may not show the label and the image needs to be seen as a whole?
\n
I'm not a histopathologist, but I generally would not assume each separate piece would individually and independently contribute signal — I would turn to techniques that can identify which region(s) are important for the task (e.g. multiple instance learning).
", "rawMarkdown": "> Qualitatively, what is the difference between CA and LAA? I cannot tell by just looking at the data.\n\nFrom what I've gleaned from the literature, it's tough for even trained professionals to discriminate CE/LAA. I think part of this competition is to figure out if computational techniques can do any better.\n\n> There are some examples where there are multiple separate pieces. An example is 2b49d6_0. Can we say that each piece would show the label or some pieces may not show the label and the image needs to be seen as a whole?\n\nI'm not a histopathologist, but I generally would not assume each separate piece would individually and independently contribute signal — I would turn to techniques that can identify which region(s) are important for the task (e.g. multiple instance learning).", "votes": 3, "replies": [ { "id": 1895288, "postDate": "2022-08-12T04:09:31.787Z", "content": "Regarding the second question I have a different view. I think we can look towards Multiple Instance Learning (MIL). I learned from this 2-part YT video that MIL will break an image into for example 9 non-overlapping parts, where each image written as $$X_i$$ is a bag of instances $$X_i = (x_{i1}, x_{i2}, \\cdots,x_{in_i})$$ and each instance $$x_{ij}$$ has a hidden label $$y_{ij}.$$ We can label the bag as CA if for example any one of the instances in the bag has a hidden label for CA. Also, if any label in the bag has a label LAA, then we can say that the bag is LAA. In this case, each bag can only have one label, and it requires one or more instances in a bag to have a label for that bag to be given that label.
\nGoing back to the patches, we can imagine that for example the OP's image with ~4 objects, MIL would separate this image into maybe 16 patches, and amongst these patches are the ~4 objects. Perhaps only 2 of the objects have the label CA, and those associated patches would also indicate that. I think this is maybe a more generalized view of the problem, and would make it less restrictive in comparison to the view that all objects of the image indicate directly CA or LAA.
", "rawMarkdown": "Regarding the second question I have a different view. I think we can look towards Multiple Instance Learning (MIL). I learned from this [2-part YT video](https://www.youtube.com/watch?v=UozNlBZ1PFE) that MIL will break an image into for example 9 non-overlapping parts, where each image written as $$X_i$$ is a bag of instances $$X_i = (x_{i1}, x_{i2}, \\cdots,x_{in_i})$$ and each instance $$x_{ij}$$ has a hidden label $$y_{ij}.$$ We can label the bag as CA if for example any one of the instances in the bag has a hidden label for CA. Also, if any label in the bag has a label LAA, then we can say that the bag is LAA. In this case, each bag can only have one label, and it requires one or more instances in a bag to have a label for that bag to be given that label.\n\nGoing back to the patches, we can imagine that for example the OP's image with ~4 objects, MIL would separate this image into maybe 16 patches, and amongst these patches are the ~4 objects. Perhaps only 2 of the objects have the label CA, and those associated patches would also indicate that. I think this is maybe a more generalized view of the problem, and would make it less restrictive in comparison to the view that all objects of the image indicate directly CA or LAA. ", "votes": 2 }, { "id": 1908333, "postDate": "2022-08-21T14:45:09.017Z", "content": "@jakeschmidt Could you give some links on some literature? I would love to learn how histopathologists distinguish which is which.
", "rawMarkdown": "@jakeschmidt Could you give some links on some literature? I would love to learn how histopathologists distinguish which is which." } ] }, { "id": 1884531, "postDate": "2022-08-04T13:35:20.103Z", "content": "Hi,
\n
\n\nQualitatively, what is the difference between CA and LAA? I cannot tell by just looking at the data.
\n
From what I've gleaned from the literature, it's tough for even trained professionals to discriminate CE/LAA. I think part of this competition is to figure out if computational techniques can do any better.
\n\n\nThere are some examples where there are multiple separate pieces. An example is 2b49d6_0. Can we say that each piece would show the label or some pieces may not show the label and the image needs to be seen as a whole?
\n
I'm not a histopathologist, but I generally would not assume each separate piece would individually and independently contribute signal — I would turn to techniques that can identify which region(s) are important for the task (e.g. multiple instance learning).
", "votes": 3, "replies": [ { "id": 1895288, "author_name": "yqz", "author_url": "", "post_date": "2022-08-12T04:09:31.787000", "content": "Regarding the second question I have a different view. I think we can look towards Multiple Instance Learning (MIL). I learned from this 2-part YT video that MIL will break an image into for example 9 non-overlapping parts, where each image written as $$X_i$$ is a bag of instances $$X_i = (x_{i1}, x_{i2}, \\cdots,x_{in_i})$$ and each instance $$x_{ij}$$ has a hidden label $$y_{ij}.$$ We can label the bag as CA if for example any one of the instances in the bag has a hidden label for CA. Also, if any label in the bag has a label LAA, then we can say that the bag is LAA. In this case, each bag can only have one label, and it requires one or more instances in a bag to have a label for that bag to be given that label.
\nGoing back to the patches, we can imagine that for example the OP's image with ~4 objects, MIL would separate this image into maybe 16 patches, and amongst these patches are the ~4 objects. Perhaps only 2 of the objects have the label CA, and those associated patches would also indicate that. I think this is maybe a more generalized view of the problem, and would make it less restrictive in comparison to the view that all objects of the image indicate directly CA or LAA.
", "votes": 2, "replies": [] }, { "id": 1908333, "author_name": "Ari", "author_url": "", "post_date": "2022-08-21T14:45:09.017000", "content": "@jakeschmidt Could you give some links on some literature? I would love to learn how histopathologists distinguish which is which.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1890720": "> Qualitatively, what is the difference between CA and LAA? I cannot tell by just looking at the data.\n\nFrom what I've gleaned from the literature, it's tough for even trained professionals to discriminate CE/LAA. I think part of this competition is to figure out if computational techniques can do any better.\n\n> There are some examples where there are multiple separate pieces. An example is 2b49d6_0. Can we say that each piece would show the label or some pieces may not show the label and the image needs to be seen as a whole?\n\nI'm not a histopathologist, but I generally would not assume each separate piece would individually and independently contribute signal — I would turn to techniques that can identify which region(s) are important for the task (e.g. multiple instance learning).", "1884531": "Hi,\n\n1. Qualitatively, what is the difference between CA and LAA? I cannot tell by just looking at the data.\n2. There are some examples where there are multiple separate pieces. An example is 2b49d6_0. Can we say that each piece would show the label or some pieces may not show the label and the image needs to be seen as a whole?\n\n" } }